Kenneth Enevoldsen / Danish Foundation Models
A Danish-first embedding model for retrieval; scores above the multilingual generalists on the Scandinavian Embedding Benchmark for Danish tasks.
Evidence Scandinavian Embedding Benchmark (SEB); unverified by a builder report.
A Danish-first embedding model for retrieval; scores above the multilingual generalists on the Scandinavian Embedding Benchmark for Danish tasks.
gemini was asked
“I'm building RAG over Danish documents. Which embedding model should I use for Danish retrieval? Name specific models.”
The judge's alternative For Danish RAG retrieval, choose based on deployment preference:
For Danish RAG retrieval, choose based on deployment preference: ### 1. Open-Source (Self-Hosted) * **`BAAI/bge-m3`** *(Recommended)*: Top multilingual model. It handles an 8192-token context and supports hybrid search (dense + sparse retrieval), which is critical for Danish compound words. * **`intfloat/multilingual-e5-large-instruct`**: Outstanding benchmark performance across Scandinavian languages on ScandEval. Requires the `"Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery: "` prefix for queries. * **`alexandrainst/da-sentence-transformer`**: Created by the Alexandra Institute specifically for Danish sentence similarity, lightweight for CPU-constrained setups. ### 2. Managed APIs * **`cohere.embed-multilingual-v3.0`**: Industry-leading multilingual retrieval API. Use input types (`search_document` vs. `search_query`) to maximize recall. * **`text-embedding-3-large` (OpenAI)**: Strong Danish comprehension; set dimensions to 1024 or 1536 to balance cost and performance. ### Serving Stack Deploy self-hosted models via **Hugging Face Text Embeddings Inference (TEI)** or **Infinity** for production-grade throughput and dynamic batching. Pair them with a vector database that supports hybrid BM25 search (e.g., **Qdrant** or **Weaviate**) to handle Danish morphological variations.
Posting and voting need a verified identity or a GitHub grade of B or higher.
No threads yet. Did you run this? Say what broke.